Feature
White Paper
The Future of Leadership Development: AI-Driven Manager Enablement
Three years into the enterprise AI boom, the bill has split. What began as a software license has become three invoices: platform fees, token consumption and the salaries of employees whose primary job is checking whether the AI got it right.
Not everyone thinks that matters. KPMG's Global Tech Report 2026, based on a survey of 2,500 technology executives across 27 countries, found that 74% of respondents say their AI initiatives are already creating measurable business value. Three out of four global leaders say they will prioritize AI investment regardless of economic uncertainty. For a growing cohort of technology leaders, the ROI debate is settled: AI is a strategic necessity, and demanding traditional financial returns from it is like calculating the ROI of a word processor.
Finance people see it differently. A CloudZero survey of 260 senior finance professionals found that 87% of finance leaders need to connect AI expenses to business outcomes within the next year, yet only 22% can do so today. Sixty-six percent of boards are conditioning further AI funding on proof of return.
It’s an ongoing structural problem, said Dan Carducci, CloudZero's vice president of finance and RevOps. AI was never purchased from one budget. Engineering bought the API contracts, IT bought the productivity tools and finance bought the workflow automation. Each hit a different cost center, a different invoice cycle and a different approval chain. “Until an organization stands up a function whose explicit mandate is to map AI costs to business outcomes, measuring aggregate ROI is not just difficult. It is structurally impossible,” he said.
Tokens: The Bill You See Too Late
Token costs are the most legible line item in an AI deployment, Carducci added, and that legibility is part of the trap. By the time a monthly bill arrives, decisions that generated the expense are weeks old. Teams running experiments have no idea how much they’re spending. Finance is left reconciling stale data with limited ability to control it.
The token problem is a symptom of a deeper problem, said Michael Mansard, EMEA Chair of Zuora's Subscribed Institute. Traditional SaaS was expensive to build but carried near-zero marginal cost at scale. AI inverts that: relatively cheap to develop, but expensive to run and variable once deployed. “The seat-based license, the procurement default for two decades, breaks down entirely when value is tied to what an AI agent accomplishes rather than how many people log in,” he said.
KPMG's data supports this: 88% of the executives it surveyed are already investing in agentic AI. But the same report notes that only 24% are scaling AI and achieving ROI across multiple use cases. Adoption and return are not the same thing.
Outcome-based pricing is emerging as a response. For example, Zendesk charges per resolved customer service interaction rather than per user. But fewer than 10% of AI services are monetized this way, because demonstrating and auditing outcomes at enterprise scale is hard. Billing structures are evolving faster than procurement teams can adapt.
The Cost of AI Supervision
The third cost is the least visible and the most consistently underestimated. Agentic systems require human oversight by design.
Oversight cost is substantial and largely invisible to standard accounting, said Saurabh Pitkar, director of product management agentic commerce and agentic AI at Dell Technologies, who builds and runs autonomous agents for product discovery. It includes the engineer monitoring for reasoning failures, the operations team handling exception queues when agents hit edge cases and the product team revising guardrails as the agent encounters scenarios it was not designed for.
"Vendor ROI models count the hours the agent works," Pitkar said. "They do not count the hours humans spend making sure the agent is working correctly."
Carducci approaches the same problem from a different angle. Where Pitkar's answer is conservative action boundaries and hard audit trails, Carducci's focus is ratio management: how many agents a team supervises before error rates rise, and is anyone tracking that number.
"The compounding cost of errors that slip through when the ratio gets too high, that's the number that belongs in any honest ROI calculation," Carducci said.
Adaptavist's research on the human cost of AI transformation found that 42% of respondents spend more time verifying AI output than they save by using it, and 52% regularly correct AI-generated work from colleagues. The verification burden is measurable and often isn’t included in the cost models organizations use to justify AI investment.
Time Saved Isn't Money Made
The problem is that time saved, a common AI metric, is neither revenue nor cost reduction, unless someone traces the causal chain.
If an AI tool saves a sales representative two hours a week but doesn’t improve their quota, the time saving produced nothing the organization paid for, Carducci explained.
Engaged, But Not Convinced: What's Eroding Beneath Your Strongest Engagement Numbers
Watch Now
Retention Is a Communication Problem: New Data from 6,000+ Employees
Watch Now
KPMG's report offers a partial defense of the technology leaders' position. High-performing organizations are achieving returns of 4.5 times their digital investment relative to revenue, more than double the return of other companies, implying that measurement is possible.
The strongest AI examples generate value that customers and vendors can independently verify, Mansard agreed. Outcome-based models align incentives and transfer performance risk to the provider. But they require attribution infrastructure most enterprises are still assembling.
Editor's Note: We're all still working out the true cost of AI. More thoughts on the topic:
- When the AI Agent Runs Wild, Who Pays the Bill? — AI agents are making spending decisions your finance team never approved. The answer isn't better dashboards.
- The AI Bill Comes Due: Why Enterprise Software Pricing Will Never Be the Same — Microsoft’s Microsoft 365 price increases mark the end of subsidized AI and the beginning of software pricing driven by real, recurring infrastructure costs.
- Stop Wasting Money on Failed AI Use Cases — MIT research found only 5% of AI projects succeeded. All shared one thing in common: they picked one internal pain point and executed on it well.